A shield construction management method and device based on digital twinning

Abstract

The application provides a kind of based on digital twinning shield construction management method and device, related to shield construction management technical field, including obtaining the basic data of karst stratum shield construction;Based on the basic data generation space-time multi-modal hypergraph;The space-time multi-modal hypergraph is fused by structure maintaining graph embedding algorithm and diffusion process based graph neural network, to obtain dynamically coupled digital twinning model;The digital twinning model is fused by persistent spectral enhancement processing and random walk algorithm, to obtain the dynamic evolution of the topological spectrum sequence of excavation risk;Based on the topological spectrum sequence of excavation risk, an adaptive excavation control strategy is generated;Based on the adaptive excavation control strategy, real-time stability evaluation is performed to obtain a dynamic management scheme for the entire shield construction process.The application can adapt to different construction stages of complex situation, improve the safety and construction efficiency in the process of shield construction.

Description

Technical Field

[0001] This invention relates to the field of tunnel boring machine (TBM) construction management technology, and more specifically, to a TBM construction management method and apparatus based on digital twins. Background Technology

[0002] In current subway tunnel engineering, shield tunneling has become the primary method for traversing complex geological conditions. However, when facing overlapping strata such as fault fracture zones, highly aquifers, and isolated boulder layers, ground disturbances, structural deformations, and the risk of instability and water inrush during tunneling are often coupled, exhibiting strong nonlinearity and high uncertainty. Existing technologies typically employ feedback-based control methods linked to construction experience rules and on-site monitoring, addressing risks through measures such as setting early warning thresholds and manually adjusting parameters. While these methods are effective in some homogeneous strata, they exhibit significant lag when facing sudden risks driven by coupled multiple fields, lacking the ability to predict and control the risk evolution path, and making it difficult to achieve dynamic, feedforward management of the entire shield tunneling process.

[0003] Therefore, there is an urgent need for a shield tunneling construction management method and device based on digital twins to solve the above-mentioned technical problems. Summary of the Invention

[0004] The purpose of this invention is to provide a shield tunneling construction management method and device based on digital twins to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0005] Firstly, this application provides a shield tunneling construction management method based on digital twins, including:

[0006] Obtain basic data for shield tunneling construction in karst formations;

[0007] The basic data is processed by a fusion of multi-scale curvature spectrum clustering method and cross-domain spectral graph generation method to obtain a spatiotemporal multimodal hypergraph that synchronously represents the multi-source relationship of stratigraphy, hydrology and mechanism.

[0008] The spatiotemporal multimodal hypergraph is fused with a structure-preserving graph embedding algorithm and a diffusion-based graph neural network to obtain a dynamically coupled digital twin model.

[0009] The digital twin model is then processed by a combination of persistent spectrum enhancement and random walk algorithm to obtain a dynamically evolving tunneling risk topology spectrum sequence.

[0010] An adaptive tunneling control strategy is generated based on the adaptive regulation of shield tunneling speed, support pressure and mud parameters according to the tunneling risk topology spectrum sequence.

[0011] Based on the aforementioned adaptive tunneling control strategy, a real-time stability assessment is performed, and the digital twin model is incrementally updated based on the real-time assessment results to obtain a dynamic management scheme for the entire shield tunneling construction process.

[0012] Secondly, this application also provides a shield tunneling construction management device based on digital twins, comprising:

[0013] The acquisition unit is used to acquire basic data for shield tunneling construction in karst formations.

[0014] The first processing unit is used to process the basic data through a fusion of a multi-scale curvature spectrum clustering method and a cross-domain spectral graph generation method to obtain a spatiotemporal multimodal hypergraph that synchronously represents the multi-source relationship between stratigraphy, hydrology, and mechanism.

[0015] The second processing unit is used to fuse the spatiotemporal multimodal hypergraph with a structure-preserving graph embedding algorithm and a graph neural network based on a diffusion process to obtain a dynamically coupled digital twin model.

[0016] The third processing unit is used to process the digital twin model by combining persistent spectrum enhancement processing and random walk algorithm to obtain a dynamically evolving tunneling risk topology spectrum sequence.

[0017] The control unit is used to adaptively control the shield tunneling speed, support pressure and mud parameters based on the tunneling risk topology spectrum sequence, and generate an adaptive tunneling control strategy.

[0018] The evaluation unit is used to perform real-time stability evaluation based on the adaptive tunneling control strategy, and to incrementally update the digital twin model based on the real-time evaluation results, so as to obtain a dynamic management scheme for the entire shield tunneling construction process.

[0019] The beneficial effects of this invention are as follows:

[0020] This invention integrates multi-source spatiotemporal data and employs innovative algorithms and mathematical models to dynamically simulate and optimize the tunnel boring machine (TBM) construction process. Specifically, based on real-time geological, hydrological, and TBM operating parameters collected from the construction site, and combined with deep learning and graph algorithms, a highly accurate and adaptive digital twin model is constructed. This model can assess the stability of the tunnel face in real time and achieve dynamic adjustments during the tunneling process through a feedforward control strategy. By incrementally updating the digital twin model, it can adapt to the complex conditions at different construction stages and adjust control variables such as tunneling speed, support pressure, and mud density in a timely manner, greatly improving the safety and efficiency of TBM construction. This method overcomes the limitations of traditional technologies in dealing with complex geological environments and sudden risks, providing a feasible solution for the intelligent and automated construction of subway TBMs.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the shield tunneling construction management method based on digital twins as described in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the shield tunneling construction management device based on digital twin as described in an embodiment of the present invention.

[0025] In the diagram: 701, acquisition unit; 702, first processing unit; 703, second processing unit; 704, third processing unit; 705, control unit; 706, evaluation unit. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0027] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0028] Example 1:

[0029] This embodiment provides a shield tunneling construction management method based on digital twins.

[0030] See Figure 1 The figure shows that the method includes steps S1, S2, S3, S4, S5 and S6.

[0031] Step S1: Obtain basic data for shield tunneling construction in karst formations;

[0032] It is understandable that this step involves shield tunneling under karst geological conditions, an engineering environment characterized by high complexity and dynamic variability, especially when traversing areas with well-developed dissolution fissures, dense accumulation of boulders, abundant groundwater, and drastic water pressure fluctuations. The core of this step lies in acquiring a comprehensive raw dataset that can depict the three-dimensional coupled system of "strata-hydrology-shield response." This raw data includes multi-temporal three-dimensional ground-penetrating radar image data, in-situ strata micro-motion response signals (such as array acoustic emission and microseismic fluctuations), real-time records of groundwater seepage pressure, tunneling control parameters during shield construction (such as cutterhead torque, thrust, and synchronous grouting volume), and displacement time-series data from key structural deformation monitoring points. By comprehensively acquiring this data, which exhibits multi-source heterogeneity, spatiotemporal non-uniformity, and response time delay characteristics, a multi-modal, deep-structure, and high-dimensional perceptual foundation is provided for the subsequent construction of a self-evolving digital twin model. Especially in karst formations, geological disturbances often exhibit latent development and abrupt changes. Collecting the aforementioned data can help capture the precursory characteristics between tunneling disturbances and geological responses in advance, enabling "holographic observation" of the dynamic evolution state.

[0033] Step S2: The basic data is processed by a multi-scale curvature spectrum clustering method and a cross-domain spectral graph generation method to obtain a spatiotemporal multimodal hypergraph that synchronously represents the multi-source relationship between stratigraphy, hydrology and mechanism.

[0034] It is understandable that this step, by creatively introducing a curvature-driven and spectral resonance fusion mechanism, maps the deep structure and dynamic response mechanism in the original heterogeneous data to a resolvable high-order graph structure. This overcomes the limitations of traditional methods that use static networks or simplified coupling models to characterize the stratigraphic-construction relationship, laying a highly malleable and structurally stable foundation for the subsequent construction of digital twin models based on graph reasoning and neural mapping. In this step, step S2 includes steps S21, S22, S23, and S24.

[0035] Step S21: Perform multi-scale curvature spectrum clustering on the basic data, and adaptively divide the curvature islands according to the fractal dimension to obtain the primary cluster of stratigraphic geometry;

[0036] Understandably, this step employs a combination of multi-scale curvature spectrum clustering and fractal dimension-driven spatial region partitioning to reveal the essential morphological characteristics of irregular geometric structures in karst strata. Specifically, firstly, for the 3D ground-penetrating radar cross-sections, fault point clouds, and in-situ displacement field configurations in the basic data, higher-order geometric curvature information (Gaussian curvature is used in this step) is calculated at different scales to construct a curvature spectral density function to characterize the local spatial change rate of geological structures. Subsequently, a multi-scale spectral clustering algorithm is introduced to perform modal decomposition of the curvature spectrum, forming preliminary spatial curvature response clusters. By selecting structural change regions corresponding to significant peaks in the spectral space, complex structures such as isolated boulders, karst cave boundaries, and fracture intersections in the strata are aggregated into potential "geometric units."

[0037] Building upon this foundation, to address the issues of ambiguous stratigraphic boundaries and multi-layered nesting in karst areas, a fractal dimension algorithm is further introduced to adaptively assess the complexity and self-similarity of curvature cluster boundaries, dynamically dividing "curvature islands" as primary stratigraphic structural units. This method fully leverages the multi-scale characteristics inherent in the structural complexity of karst strata, thereby avoiding information fragmentation caused by manual or rule-based partitioning. Ultimately, a set of primary stratigraphic geometric clusters with high internal consistency, clear boundaries, and usability for subsequent cross-domain map modeling is obtained.

[0038] Step S22: Perform cross-domain spectral resonance alignment processing on the primary cluster of stratigraphic geometry and the seasonal groundwater level time series in the basic data to obtain a multi-source spectral resonance matrix;

[0039] Understandably, this step first performs graph embedding processing on the spatial morphology within each "curvature island" in the primary cluster of stratigraphic geometry to extract the dominant frequency component representing the micromorphological evolution of the region; at the same time, it uses Empirical Mode Decomposition (EMD) combined with Hilbert transform to perform non-stationary feature decomposition on the groundwater level time series data to obtain the intrinsic mode functions and corresponding instantaneous spectra that vary with time.

[0040] Subsequently, regions exhibiting resonant coupling between geometric evolution patterns and hydrological change frequency bands were searched in the spectral space. To address the inconsistency between the timescales of formation response and hydrological change, a non-uniform Fourier transform algorithm was introduced to reconstruct the spectral frequency distribution, enabling alignment and matching of the curvature change spectrum and the water level change spectrum across multiple frequency bands. The coupling strength between the two types of spectra was calculated using a resonance exponential function, constructing a multi-source spectral resonance matrix with "curvature island region-hydrological frequency band" as node pairs. The values ​​in the matrix reflect the resonant synchronicity between formation units and corresponding hydrological features.

[0041] In this step, the resonance exponential function is as follows:

[0042] RIF(t)=α·H(t)+β·C(t)+γ·P(t)+δ·J(t)+ζ·E(t)

[0043]

[0044] C(t) = exp( - ∥R model (t)-R criteria (t)∥ 2 )

[0045]

[0046] Wherein, RIF(t) represents the resonance exponential function value, reflecting the intensity of the system's resonance response to the control strategy at the current moment; t represents the current time; α, β, γ, δ and ζ are weighting coefficients, representing the relative importance of each item in the total resonance exponent, which can be set by fitting or experts; H(t) is the peak disturbance sensitivity exponent, which measures the degree of maximum response concentration in the sensitivity tensor at the current moment. R represents the sensitivity tensor at location x and time t, Ω represents the domain of the formation or spatial region; C(t) is the temporal consistency coefficient, which evaluates the degree of consistency between the model output and the criterion map in terms of trend; model (t) represents the stability response vector output by the digital twin model, R criteria (t) is the stability reference vector in the criterion spectrum; P(t) is the spectral resonance energy density, which reflects the proportion of energy in the disturbance response that coincides with the formation resonance frequency; In response to the spectral components at frequency λ (obtained via Laplace spectral decomposition), Λ * The set of resonant frequencies, i.e., the spectral band containing the intrinsic frequencies of the strata or structure; J(t) is the jump state frequency, the number of violent state transitions (jumps) occurring per unit time; s i →s j For state transition events, from state s i Transition to state s j ∈ represents the threshold for state jumps; exceeding this value is considered a sudden state. Δt represents the size of the time window. E(t) is the regional risk potential gradient, reflecting the maximum risk potential change value within a spatial range at a certain moment. Φ(x,t) represents the risk potential value at position x, which is usually extracted from the potential matrix. Φ(x',t) represents the risk potential value at position x'.

[0047] Step S23: Perform cross-domain spectral graph generation processing on the multi-source spectral resonance matrix to obtain a priori supernode set containing three types of supernodes and their weights: virtual cave, water mass, and fault.

[0048] Understandably, this step begins by introducing Graph Matching Networks (GMNs). This algorithm, through cyclically exchanging node representations and structural attention weights between source-target graph pairs, extracts implicit isomorphic substructures and highly similar subgraph patterns from the stratigraphic spectrum, hydrological spectrum, and stress perturbation spectrum, thereby inferring the potential mapping relationships across spectral structures. With this cross-domain structural transfer, the model can identify in the frequency domain "where abrupt changes in stratigraphic morphology will be synchronized with which type of high-frequency hydrological band or stress spectrum jump," thus aggregating similar dynamic behaviors originally scattered across their respective spectra into a unified similarity-weighted graph. This provides high-confidence structural edge weights and spatial location constraints for the subsequent generation of virtual nodes.

[0049] After obtaining spectral domain structure matching information, a node embedding method based on spectral graph wavelet (SGWNE) is further employed to perform multi-scale wavelet diffusion on local high-energy subdomains in the resonance matrix. By propagating local perturbations within the Graph Laplacian eigenspace, SGWNE simultaneously captures three-dimensional features of frequency response, resonance amplitude, and adjacency topology to synthesize three types of virtual supernodes: ① virtual cave nodes, representing potential geological cavity centers corresponding to abrupt changes in the cavity spectrum; ② water mass nodes, characterizing high-permeability, water-rich areas formed by the aggregation of hydrological perturbation spectra; ③ fault nodes, mapping tectonic fracture zones implied by high-gradient stress spectral bands. The edge weights between supernodes utilize the cross-domain similarity coefficients learned by GMNs to form a priori set of supernodes, which is then embedded as a highly abstract "risk gene" into the subsequent spatiotemporal multimodal hypergraph, providing a physically consistent and evolvable structural starting point for the digital twin model.

[0050] This step breaks away from the traditional approach that relies on physical drilling or ground-penetrating radar. By using cross-domain learning and structure generation based on frequency domain spectral resonance features, it constructs a virtual high-risk node cluster coupled with geological, hydrological, and stress domains. This not only improves the physical representation integrity of the hypergraph structurally but also significantly enhances the subsequent digital twin model's ability to predict and respond to unobservable instability factors.

[0051] Step S24: Perform spatiotemporal supergraph reconstruction processing on the prior supernode set to obtain a spatiotemporal multimodal supergraph that synchronously represents the multi-source relationship between strata, hydrology, and mechanism.

[0052] Understandably, in this step, a Dynamic Hypergraph Neural Network (DHGNN) is first used to learn the high-order structure of the prior hypernode set: three types of nodes—virtual caves, water masses, and faults—are embedded into a unified spectral space. The spectral resonance intensity and spatial proximity between nodes are used as constraints to iteratively update the weights and order of the hyperedges, enabling caves, water masses, and faults to establish many-to-many relationships within the same hyperedge. Subsequently, a Time Graph Attention Network (TGAT) is introduced to temporally encode real-time construction parameters such as tunneling speed, support pressure, and grouting volume. These dynamic control variables are injected into the hypergraph as temporal features, and the coupling strength between nodes in different time slices is characterized by attention weights. Thus, the time dimension is superimposed on the original spatial hyperedge to obtain a spatiotemporal multimodal hypergraph that simultaneously contains spatial topology and temporal dependencies.

[0053] By combining high-order correlation modeling of DHGNN with temporal attention fusion of TGAT, this step enables the supergraph to capture the nonlinear coupling laws between the three domains of "stratum morphology-hydrological fluctuation-stress migration" at the structural level, and to update the weights of the shield tunneling control behavior in real time at the temporal level, thus realizing a refined and dynamic expression of the complex karst subway tunnel environment. The technical effect is that it significantly reduces the structural entropy during multi-source information fusion, avoids the information loss and lag problems of traditional single-domain graph models, and provides a high-fidelity computational graph structure for subsequent digital twin prediction.

[0054] This step transforms static prior knowledge into a dynamic structural expression with time-evolution capabilities. This not only more accurately reflects the collaborative evolution characteristics among multiple factors during shield tunneling, but also lays a graph structure foundation for subsequent digital twin structure-based mapping and coupling modeling. This enhances the interpretability and real-time reasoning capabilities for risk areas and mechanisms in complex karst subway tunnel scenarios.

[0055] Step S3: The spatiotemporal multimodal hypergraph is fused with a structure-preserving graph embedding algorithm and a graph neural network based on diffusion process to obtain a dynamically coupled digital twin model;

[0056] It is understandable that this step ensures that the complex connection relationship between multi-source heterogeneous nodes is not distorted in low-dimensional space by maintaining the embedded structure. At the same time, the changes in shield tunneling construction parameters and the response of the stratum-hydrological system are dynamically injected into the twin structure in a physically consistent manner, so that the constructed digital twin model has both topological stability and dynamic response capability. In this step, step S3 includes steps S31, S32, S33 and S34.

[0057] Step S31: Perform high-order structure embedding processing based on the spatiotemporal multimodal hypergraph to construct a structure-preserving embedding vector set. The structure-preserving embedding vector set is a vector set that preserves the formation fracture mode, grouting behavior and seepage pressure disturbance in the high-dimensional embedding space.

[0058] Understandably, this step first employs a Hypergraph Neural Network (HGNN) to perform joint feature propagation on the hyperedge relationships formed by multiple nodes (such as faults, water masses, and grouting points) in the multimodal hypergraph, capturing the high-order semantic dependencies between them. HGNN can effectively model "many-to-many" entity interactions, and in this scenario, it can be used to express the "linkage effect of grouting behavior in a certain area on multiple fracture nodes and their surrounding water pressure fields." Such high-order relationships are difficult to identify by traditional graph embedding models.

[0059] Subsequently, to preserve the structural similarity and evolutionary synergy among nodes during the embedding process, this step introduces a second-order structure-preserving embedding strategy. This strategy emphasizes not only preserving the first-order adjacency relationships of nodes (such as "a water mass node is connected to its neighboring fault"), but also preserving the consistency of its adjacency environment with other similar nodes in the network structure (i.e., "the common response pattern of the water mass and the fault"). During processing, grouting behavior is transformed into dynamic edge attributes with time-series labels, and seepage disturbance is modeled as a time-varying weight vector, which is then projected into the embedding space along with the fracture structure embedding vector. Through this fusion mechanism, the model obtains a set of structure-preserving embedding vectors, which not only express the distribution characteristics of the geological structure but also imply the dynamic coupling of construction disturbance and hydrological response in time and space.

[0060] Step S32: Perform nonlinear isomorphic transformation processing based on the structure-preserving embedded vector set, and perform cross-domain mapping between the structural disturbance trajectory in the stratum attribute space and the feedback evolution path in the shield control variable space to construct an isomorphic nonlinear mapping model.

[0061] Understandably, this step first employs the t-distributed random neighborhood embedding algorithm (t-SNE) to perform nonlinear dimensionality reduction on the "fracture mode", "grouting behavior sequence" and "seepage disturbance path" in the embedded vector set, maintaining the neighborhood structure of these stratum attributes in the transformed space as similar to that in the original space. Then, combined with manifold alignment technology, the embedded stratum attribute space is paired with the shield control parameter space (such as the feedback evolution sequence of propulsion speed and cutterhead torque) to establish the correspondence between the two spaces.

[0062] In the processing, the formation disturbance trajectory is treated as a set of topological deformation manifolds evolving over time, while the sequence of shield tunneling control variables serves as its potential excitation source. This cross-domain mapping process applies isostructural constraints (i.e., preserving the distance relationships within each space) to the low-dimensional embedding structure of the two modes and establishes a mapping function between them, enabling the model to learn cooperative patterns across spaces while maintaining the local structure. The resulting isostructural nonlinear mapping model has the ability to predict formation response paths from the control space and also supports inferring possible changes in control inputs from the disturbance trajectory, forming a two-way inference closed loop.

[0063] This step breaks down the barriers between independent modeling of stratigraphic information and shield tunneling control data. By using nonlinear manifold alignment technology, it achieves deep integration of the two types of data at the levels of structural preservation and semantic interpretation, significantly improving the model's ability to perceive the dynamic causal relationship of "control-response". It is particularly suitable for the problem of geological feedback lag and fuzzy intertwining of disturbance evolution in karst strata, and provides a unified expression framework for subsequent causal diffusion modeling and strategy prediction of digital twin models.

[0064] Step S33: Based on the isomorphic nonlinear mapping model, perform modeling processing of the graph neural network based on the diffusion process to construct a graph neural network that reflects the relationship between the stratum disturbance response mechanism and the shield control behavior.

[0065] Understandably, this step builds upon the isomorphic nonlinear mapping model constructed in the previous stage, introducing a graph neural network modeling method based on diffusion processes (D-GNN) to simulate how construction disturbances propagate along complex geological structures and, in turn, affect the adjustment of tunnel boring machine parameters. First, the control variables embedded in the isomorphic mapping model and the stratigraphic attribute nodes are projected together into a unified graph structure, and a graph diffusion kernel is defined based on the spatial adjacency relationships between nodes and the disturbance evolution path. This diffusion process employs a diffusion convolutional neural network, the core idea of ​​which is to use a probability transition matrix on the graph structure to simulate the propagation probability of disturbances from one node to another, thereby updating the node states. This network is better able to capture spatial dynamic changes and is particularly suitable for nonlinear conduction phenomena in karst strata, such as groundwater inrush and surrounding rock shear failure.

[0066] Simultaneously, to enhance the model's ability to express physical mechanisms, this step also integrates a physical-guided graph diffusion mechanism, embedding mechanical control equations such as the Darcy seepage equation and stress-strain relationship as edge constraints into the diffusion kernel. Thus, each diffusion hop is not only a result of information propagation but is also controlled by physical properties such as formation permeability and shear modulus, giving the diffusion process realistic mechanical constraints. This simulates the dynamic mechanism of "adjustment of tunneling force triggering changes in grouting pressure, leading to the migration of water pressure towards the fault channel." The resulting graph neural network based on the diffusion process possesses comprehensive capabilities for propagating disturbances along geological topology, evaluating response intensity, and providing feedback for regulatory behavior.

[0067] Step S34: Construct a dynamic twin mechanism based on the graph neural network based on the diffusion process. By performing dynamic graph memory modeling on the diffusion structure of the graph neural network based on the diffusion process at the time series level, a dynamically coupled digital twin model is constructed.

[0068] Understandably, this step revolves around constructing a dynamic twin mechanism based on a graph neural network structure for the diffusion process. The core lies in introducing time-dimensional state evolution modeling, enabling each node's state (such as seepage pressure, stress, grouting intensity, etc.) to form a traceable and predictable dynamic trajectory as the tunnel boring machine advances. First, a dynamic graph neural network framework is used to model the continuous evolution of the diffusion graph over time. Here, a gated recursive mechanism is introduced to dynamically update the graph convolution weights, allowing the network to automatically adjust its diffusion mode at each time step to reflect real-time changes such as ground softening, water pressure rebound, and control quantity adjustments.

[0069] To more accurately preserve the impact of historical states on current disturbance trends, this step further incorporates a graph memory mechanism for dynamic graph memory modeling. This step employs the TGCRN (Temporal Graph Convolutional Recurrent Network) model structure to extract the historical state sequences of the diffusion graph nodes using graph convolutional gated recurrent units, constructing node-level time memory vectors. In this scenario, the fluctuation sequences of shield tunneling control parameters (such as the advance rate curve) and ground responses (such as stress release cycles) are jointly modeled through dynamic adjacency relationships in the graph structure. This ensures that each node not only encodes its current state but also inherits its disturbance history, thus forming a "time-sensitive propagation map" in the embedding space.

[0070] Ultimately, a dynamically coupled digital twin model is formed that evolves over time, dynamically adjusts the structure, and preserves the physical response. This model can simulate the impact of control behavior on the multi-field coupled system in real time during shield tunneling and provide predictive assessments and feedback inputs for each time step.

[0071] This step not only improves the model's ability to express complex construction disturbance behavior over time, but also significantly enhances the accuracy of predicting future state change trends. It is a key supporting link for subsequent risk feedforward identification, control strategy generation, and model self-updating. It is particularly suitable for high-risk tunneling environments in karst subway tunnels where control parameters fluctuate drastically and environmental responses are strongly coupled.

[0072] Step S4: The digital twin model is subjected to a fusion process of persistent spectrum enhancement and random walk algorithm to obtain a dynamically evolving tunneling risk topology spectrum sequence.

[0073] It is understandable that this step, by combining persistent spectrum enhancement processing with random walk algorithms, can extract richer and more comprehensive tunneling risk topology information from the digital twin model. This not only provides long-term structural evolution that traditional risk assessment methods cannot capture, but also generates a dynamic, time-series risk spectrum sequence at each time step, providing an accurate risk assessment basis for subsequent construction process optimization, control strategies, and risk prediction. In this step, step S4 includes steps S41, S42, S43, and S44.

[0074] Step S41: On the time frame of the digital twin model, perform spectral wavelet transform processing on the supernodes of the spatiotemporal multimodal hypergraph based on spectral graph theory, and combine the Vietoris-Rips complex construction method to gradually map the spatiotemporal connection relationship between supernodes into nested hierarchical simplex complexes to obtain a cross-scale topological filter cluster.

[0075] Understandably, this step employs a graph wavelet framework to construct a spectral filtering kernel function based on the graph Laplacian operator, performing wavelet expansion on the adjacency relationships of supernodes at different scales. This process not only identifies high-frequency disturbances in areas of local water pressure anomalies or stress concentration, but also preserves the low-frequency topological skeleton of fault boundaries or karst structures, allowing the same hypergraph to exhibit distinct topological patterns in different frequency domains. This transformation is particularly suitable for scenarios involving multiple intertwined disturbances, such as shield tunneling through karst fracture zones, effectively improving the ability to identify heterogeneous boundary regions and regions with coordinated multi-source anomalies.

[0076] Subsequently, based on the wavelet activation results, the Vietoris-Rips complex construction method is introduced to perform incremental topological modeling of the spectral response similarity relationships between nodes. By progressively expanding pairs of supernodes with mutual distances below a set threshold within each time frame, a nested topological complex containing high-dimensional simplexes such as points, edges, triangular faces, and tetrahedrons is constructed. This process supports dynamic updates: as each new set of data is acquired during tunnel boring machine (TBM) advancement, the system incrementally expands the existing complex structure without requiring full map reconstruction, thereby achieving temporal topological tracking of geological disturbance patterns and control feedback pathways. Finally, a set of cross-scale topological filter clusters is obtained, which characterize the continuity and abrupt changes of the formation response mechanism with a multi-layered complex structure, and provide a compressible and computable topological feature space for subsequent risk spectrum construction and causal inference.

[0077] Step S42: Perform persistent spectrum enhancement processing on the topological filter cluster according to the method of fusion β distribution fitting and persistent entropy measurement to obtain a set of homogeneous persistent feature vectors;

[0078] Understandably, this step first generates a persistent bar chart based on a cross-scale topological filter cluster. This chart represents the time difference between the generation and disappearance of homology classes (such as connected components, loop structures, cavities, etc.) in each dimension of the topological filter cluster, i.e., "birth-death" pairs. Subsequently, a β distribution is introduced as the fitting function for the normalized persistence of the bar chart to more accurately reflect the probability distribution characteristics of the duration of topological features. This approach is more suitable for asymmetric persistence distributions than traditional equal-weighted or normal distribution fitting, and is especially suitable for scenarios with extremely skewed homology features in karst structures (e.g., short-lived hydrostatic anomaly loops vs. long-term fault zone loops).

[0079] After fitting the β-distribution, the information uncertainty in each topological dimension is calculated using persistence entropy. This identifies key topological structures with high persistence and low entropy—such as stable cavern envelopes, persistent hydraulic migration paths, or local stress barrier boundaries—which are considered structural units highly correlated with risk evolution. Finally, these selected topological features are transformed into a unified high-dimensional numerical representation, forming a homogeneous persistence feature vector set. Each vector element corresponds to a topological structure spanning spatiotemporal scales, and its value reflects the structure's stability, coupling, and anomaly characterization ability in both the time and scale domains.

[0080] Understandably, this step, by integrating β-distribution fitting and persistent entropy measurement, achieves quantitative screening and characterization of key morphologies in multi-scale topological structures, thereby significantly improving the accuracy of digital twin models in identifying long-term evolutionary structures in spatiotemporal disturbance patterns. This method is particularly suitable for capturing stable, deep-seated risk configurations when shield tunnels traverse karst areas, fault zones, or water inrush zones, providing a highly discernible and low-redundancy structural feature foundation for subsequent construction of risk topology spectrum sequences and feedforward control strategies.

[0081] Step S43: Perform random walk trend field encoding processing based on the homology persistence feature vector set to generate a node-edge risk potential energy matrix that reflects the sensitivity of water inrush-collapse coupling;

[0082] Understandably, this step, by applying the random walk algorithm to a graph structure, aims to simulate the propagation process of the impact of potential unforeseen events such as water pressure fluctuations and ground collapses on construction activities. Specifically, the random walk algorithm assigns a potential "potential energy" value to each node (e.g., caves, faults, water masses), reflecting the node's response strength when faced with disturbances. By simulating the walk process, the algorithm evaluates the energy diffusion and feedback paths between nodes caused by geological changes (e.g., sudden water inrush, cave collapse). In particular, this approach can assign risk values ​​to the interactions between nodes and their neighboring nodes, reflecting the propagation trajectory of potential instability.

[0083] While performing random walks, potential encoding helps convert the potential energy calculation of each node and its corresponding edge into a numerical risk scale. Potential encoding employs a weighted random walk model based on a diffusion process. By calculating the potential changes between nodes and edges and combining this with their persistence characteristics (such as the spatiotemporal correlation of stability), an accurate node-edge risk potential matrix is ​​generated. This matrix contains the strength of the risk correlation between each node and its neighboring nodes at a specific time series, especially the sensitivity to common geological events such as water inrush and landslides. This process not only reveals the sensitivity of different stratigraphic units to disturbances during construction but also helps identify potentially high-risk areas and predict their evolution as construction progresses.

[0084] This step, by combining random walk and potential coding, quantifies and visualizes the spatial propagation process caused by various risk factors such as water inrush and landslides during construction. The generated node-edge risk potential matrix provides a comprehensive and dynamic risk assessment tool that can accurately identify high-risk nodes and their interactions during construction, providing strong data support for subsequent risk warning, control strategy optimization, and construction scheduling. This method is particularly suitable for real-time risk prediction and decision support in complex geological environments, improving the ability to predict construction safety under complex hydrogeological conditions.

[0085] Step S44: Perform time series clustering analysis based on the node-edge risk potential energy matrix. By capturing the migration and aggregation trajectory of risk potential energy peaks in the time domain, obtain a dynamically evolving tunneling risk topology spectrum sequence.

[0086] Understandably, this step first splits the node-edge risk potential energy matrix into multiple time slices according to the time series, constructing a risk propagation atlas with a time order. In each frame, the principal spectral components of the graph structure are extracted through graph Laplacian spectral decomposition. These components reflect the non-uniformity of risk energy distribution and propagation patterns in the graph. Subsequently, time series clustering analysis is used to track the trajectories of high-energy spectral components in different time frames, identifying which nodes and their associated edges consistently exhibit high-risk peaks across multiple time periods, forming "risk potential energy channels" in the time domain.

[0087] To enhance the coherence and robustness of convergence trajectories, a dynamic time calibration algorithm is introduced to register spectral structures across different time frames, ensuring that even if risk peaks shift over time, their convergence and identification in the spectral space can be maintained. During this process, if a structural feature (such as a water mass-fault path) is found to be consistently in a high-risk spectral band across multiple consecutive time slices, it is identified as an "evolving risk core topology," and its trajectory is marked as a spectral peak migration path.

[0088] Finally, through temporal aggregation and spectral energy integration, a dynamic tunneling risk topology spectrum sequence reflecting the risk propagation and evolution trend at each stage of the construction process is generated. This sequence expresses the position, intensity, and trend changes of potential instability factors at the tunnel face in the graph structure in the form of temporally arranged spectral feature vectors, possessing significant predictive and regulatory guidance significance.

[0089] This step, through temporal spectrum clustering, successfully compresses the spatiotemporal distribution of multi-source risk potential energy in the graph structure into an analytical spectral sequence. This not only improves the granularity of identifying complex coupled risks, but also provides quantifiable input indicators for feedforward risk early warning and dynamic path adjustment in shield tunneling construction.

[0090] Step S5: Based on the tunneling risk topology spectrum sequence, adaptive control of shield tunneling speed, support pressure and mud parameters is generated to produce an adaptive tunneling control strategy.

[0091] It is understandable that this step transforms complex, high-dimensional risk spectrum information into dynamic control actions within the parameter space. This method constructs a tunneling control mechanism that does not rely on human rules or preset thresholds, achieving a closed-loop response from risk identification to control feedback. Especially in complex strata with well-developed karst fractures and frequent hydrological changes, this strategy can effectively mitigate the amplification effect of disturbances during construction, improve the safety redundancy and operational efficiency of shield tunneling. In this step, step S5 includes steps S51, S52, S53, and S54.

[0092] Step S51: Perform counterfactual causal inference processing based on the tunneling risk topology spectrum sequence. By injecting the shield control quantity change as an exogenous do intervention into the hypergraph-Do Bayesian network, a multi-field coupled causal structure graph with scenario-based counterfactual weights is generated.

[0093] Understandably, this step first involves matching key topological features (such as local peak frequencies and amplitude trends) in the risk spectrum sequence with shield tunneling control variables (such as increased advance rate and decreased grouting density) to construct an initial structure of a multimodal causal graph. Then, a structural learning algorithm (such as the Peter & Clark PC algorithm) is used to extract potential causal edges between variables. Based on this, the control variables (shield tunneling parameters) are used as exogenous intervention nodes and injected into the graph structure through the do(X=x) operation to generate a hypergraph-Do Bayesian network. This network can simulate the evolution trend of risk spectrum features under a given virtual control input.

[0094] Unlike traditional causal inference, the causal graph structure here is not limited to single variables or first-order relationships, but is built upon a prior hypergraph model and spectral topology, exhibiting higher-order structural dependencies. Therefore, the inference process is no longer a linear path of "control variable → risk variable," but rather a multi-hop, multi-modal structural propagation process, such as: changes in propagation rate → intensified water mass fluctuations → inflection point of abnormal seepage pressure → enhanced local spectral convergence → increased risk of sudden flooding. This structure, by assigning counterfactual weights to different causal paths, can simulate evolutionary trajectories that "did not happen but may happen," thereby assessing the potential risk distribution under multiple control strategies.

[0095] This step significantly enhances the model's structural understanding of the "control change - risk evolution" mechanism and its ability to generalize scenarios by constructing a multi-field coupled causal structure diagram with counterfactual intervention capabilities. It can predict the chain risk response that may be triggered under specific construction behavior adjustments, providing a causal interpretability basis for the generation of subsequent control strategies.

[0096] Step S52: Perform fractional manifold spectrum mapping on the multi-field coupled causal structure diagram according to the Caputo fractional manifold embedding method to obtain the manifold embedding tensor of measurable shield control-stratum response dynamics.

[0097] Understandably, this first step involves the Caputo fractional manifold embedding method, based on manifold learning theory and fractional calculus, which effectively handles dynamic systems with memory effects and historical dependencies. In this process, each node (e.g., advance rate, support pressure, formation displacement, etc.) and edge (e.g., control feedback relationships between nodes, formation disturbance propagation channels) in the graph is embedded into a high-dimensional manifold space. The topology of this space reflects the complex nonlinear coupling relationship between control quantities and formation responses.

[0098] The Caputo fractional manifold embedding method operates as follows: First, the causal graph is spectrally decomposed. By analyzing the eigenvalue spectrum obtained from the graph Laplacian operator, a non-Euclidean space suitable for representing formation response dynamics is constructed. Then, the Caputo fractional derivative is used to introduce "non-integer order" variations. This means that the mapping between control variables and formation responses considers not only instantaneous responses but also the cumulative effects of historical states, allowing for the consideration of more complex memory effects and hysteresis characteristics during system response. In this way, the time dependence of nonlinear formation responses and construction controls can be captured.

[0099] Ultimately, through these fractional-order manifold spectrum mappings, the resulting manifold embedding tensor not only accurately represents the spatiotemporal coupling relationship between control behavior and formation response, but also measures the dynamic characteristics of each control variable in relation to the formation response. This tensor provides a higher-dimensional, more refined feature space for subsequent risk assessment and control strategy generation.

[0100] This step, by introducing the Caputo fractional manifold embedding method, enables the establishment of a more accurate nonlinear dynamic model between the formation and the shield tunneling control, particularly in handling memory effects and historical dependencies that are difficult to capture using traditional methods. This method can better describe the long-term dynamic evolution of the formation response when facing complex construction environments and multiple disturbances, enhancing the predictive power and spatiotemporal adaptability of the digital twin model.

[0101] Step S53: Based on the manifold embedding tensor, construct a variable step size Stackelberg difference game model on the strategy surface of tunneling speed-support pressure-slurry density, and solve the Pareto optimal response curve to obtain the staged control candidate tensor.

[0102] It is understandable that the variable-step Stackelberg differential game model constructed in this step involves, based on the Stackelberg game framework, designating the tunnel boring machine (TBM) construction control system (master control unit) as the leader, and the natural response systems such as the strata and hydrology as followers. The leader first makes control actions such as tunneling speed, and the followers then respond accordingly with stress migration, water pressure changes, etc. Unlike traditional static games, this model considers the high time correlation and feedback lag of geological disturbances, thus introducing a differential game modeling approach to express the transmission process of control and response over time in discrete time steps. To address the unstable geological feedback rate during construction, a variable-step mechanism is specifically introduced, allowing the model to automatically shorten the control adjustment step size in highly sensitive response sections (such as near karst caves or fault fracture zones), improving game convergence accuracy; while in relatively homogeneous strata sections, the time span is widened to improve computational efficiency.

[0103] In the game-theoretic solution phase, a multi-objective function is constructed with the goals of "minimizing formation disturbance response error," "controlling risk level," and "improving energy efficiency." Combining evolutionary game strategy search and Pareto front analysis, the optimal control trajectory under different regulatory constraints is solved, forming a family of Pareto optimal response curves describing the relationships between control variables. These curves exist as equipotential paths on a three-dimensional strategy surface of tunneling speed, support pressure, and mud density, reflecting the optimal regulatory distribution under current geological conditions.

[0104] Finally, these curves are divided into stages (such as by geological type, tunneling ring number, risk level, etc.) to construct a staged control candidate tensor, where each sub-tensor represents a set of control strategies for a certain stage, which has deployability and real-time selectivity.

[0105] Step S54: Solve the Hamiltonian variational inequality based on the control candidate tensor to obtain an adaptive tunneling control strategy that satisfies the dynamic stability constraint.

[0106] Understandably, this step first establishes a dynamic system model based on the control candidate tensor, which includes the nonlinear relationship between control inputs (such as tunneling speed and support pressure) and response variables (such as formation stress and seepage water pressure) during the tunneling process. Then, a Hamiltonian function is constructed, which typically includes control variables, response variables, and the system's state equations. The Hamiltonian function aims to describe the overall dynamics of the control system, including objective functions (such as risk minimization and efficiency maximization) and constraints (such as dynamic stability and energy consumption limits). In this function, the interaction between control variables and state variables is represented by gradients and Lagrange multipliers. When solving the Hamiltonian variational inequality, the problem is transformed into an optimization problem by finding an equilibrium point between the control and response variables that minimizes the change in the Hamiltonian while ensuring system stability under external disturbances (such as water pressure fluctuations and abrupt changes in formation). By solving this inequality, we can obtain an adaptive tunneling control strategy that meets dynamic stability constraints. This step utilizes the Hamiltonian variational inequality solution method to provide a systematic approach for optimizing the control strategy of complex nonlinear dynamic systems during tunnel boring machine (TBM) construction. This method not only enables optimal selection of control variables under varying geological and hydrological environments but also allows for real-time updates and adjustments to the strategy to adapt to external disturbances, thereby maintaining stability during tunneling. Ultimately, the resulting adaptive tunneling control strategy exhibits strong robustness and flexibility, significantly improving risk warning capabilities and safety assurance levels during construction, particularly demonstrating significant advantages under complex geological conditions.

[0107] Step S6: Perform real-time stability assessment based on the adaptive tunneling control strategy, and incrementally update the digital twin model based on the real-time assessment results to obtain a dynamic management scheme for the entire shield tunneling construction process.

[0108] It is understandable that this step, based on a dynamic management scheme of real-time stability assessment and incremental updates using digital twins, can significantly improve the intelligence level of the tunnel boring machine (TBM) construction process. This scheme can respond in real-time to changes in the geological and hydrological environment, automatically optimize control strategies, and improve construction safety, efficiency, and stability, making it particularly suitable for high-risk TBM construction under complex geological conditions. Through this continuous optimization and update mechanism, TBM construction management can achieve efficient, intelligent, and flexible full-process control, ensuring the smooth progress of the project in a dynamic environment. In this step, step S6 includes steps S61, S62, S63, S64, and S65.

[0109] Step S61: Based on the probabilistic graphical convolutional network and the generalized information gain analysis method, the adaptive tunneling control strategy is subjected to stability-sensitive response mapping processing to construct the nonlinear sensitivity distribution tensor between the tunneling control quantity and the disturbance response.

[0110] Understandably, this step introduces probabilistic graphical convolutional networks and generalized information gain analysis to perform stability-sensitive response mapping on the previously obtained adaptive tunneling control strategy. Specifically, tunneling control variables (such as advance speed, support pressure, and mud density) are modeled as input features of nodes in the graph, while disturbance responses such as formation, water pressure, and structural stress are set as output node responses. In the graph structure, a probabilistic modeling mechanism is introduced, making the control input not only a deterministic value but also a distribution function with uncertainty, thereby capturing the uncertain diffusion path of disturbance responses propagating in the graph structure. Based on this, the "information contribution" of each control variable to the disturbance response distribution is measured using a generalized information gain function—that is, the intensity of response entropy change or distribution drift caused by the change of the control variable in multi-physics disturbances. A multi-dimensional nonlinear sensitivity distribution tensor is constructed using the above information gain measurement results, where each tensor unit reflects the sensitivity intensity of a certain control variable to a specific type of disturbance (such as sudden pressure rise or abnormal tension in the support structure) at a certain time point and in a certain region. This sensitivity tensor not only reveals the nonlinear mapping relationship between the control variables and the disturbance response, but also provides a high-resolution data foundation for subsequent spatial stability identification and strategy optimization.

[0111] This step, by introducing a structured graph neural network and a probability gain measurement mechanism, enables the quantification of the response sensitivity distribution of control parameters to multi-physics disturbances while maintaining topological dependencies. This significantly improves the explanatory power and optimization basis of the control strategy for the geological response system, laying a key foundation for the spatiotemporal perception and risk prevention and control mechanism of digital twin models.

[0112] Step S62: Based on the spectral decomposition-modulus centroid back-inference method of the Laplace tensor domain, the spatial stability region of the disturbance response sensitivity distribution tensor is calibrated to obtain the spatiotemporal stability criterion spectrum.

[0113] It is understandable that the nonlinear sensitivity distribution tensor constructed in the previous step of this process uses the spectral decomposition-modulus centroid back-induction method in the Laplace tensor domain to perform spatial stability region calibration on the multi-field coupled response caused by shield tunneling construction control disturbance.

[0114] The Laplace tensor domain spectral decomposition-modulus centroid back-inference method first utilizes the Laplace tensor spectral decomposition method to expand the original sensitivity tensor in the frequency domain, identifying the dominant modes and spectral features that best represent the severity of the perturbation. In this process, the tensor Laplace operator of the Laplace tensor allows for simultaneous consideration of spatial location, perturbation type, and time series structural dependencies in a multidimensional space, thereby better preserving the topological patterns of the perturbation response.

[0115] Subsequently, a modulus centroid back projection method is introduced. Using the high-amplitude perturbation modes obtained from spectral decomposition as input, the modulus energy distribution in tensor space is calculated, and a perturbation "centroid map" is constructed based on the spatial clustering trend of the perturbation intensity. By backprojecting the centroid map into the geological physical model, potential highly sensitive areas and instability-inducing zones are mapped, thereby delineating the boundaries of key stability spatial regions during tunnel boring. Combining the temporal evolution dimension of the perturbation, a spatiotemporal stability criterion map is finally generated. Each regional unit not only includes the perturbation response intensity level but also carries the sensitivity trend for the corresponding time period, the corresponding interval of control variables, and the risk threshold.

[0116] This step transforms the high-dimensional nonlinear disturbance sensitivity tensor into a criterion map with spatial resolution and temporal predictive power, enabling the system to quantitatively characterize the stability distribution of tunnel boring machine (TBM) construction. This map effectively identifies "high-sensitivity-high-risk response zones," providing spatial guidance and dynamic reference for subsequent control strategy implementation and twin model updates. It significantly enhances the spatiotemporal adaptability and scenario generalization ability of the entire model system, making it particularly suitable for urban TBM engineering environments with complex geological changes and significant superposition of disturbance sources.

[0117] Step S63: Construct an adversarial residual filtering network based on consistency constraints, compare the stability response trend of the current digital twin model output with the criterion map, extract the model offset residuals while maintaining the topological invariance of the structural tensor, and obtain the incremental update vector group.

[0118] Understandably, this step constructs an adversarial residual filtering network based on consistency constraints to achieve high-precision identification and residual extraction of the response shift between the current digital twin model output and the stability spatiotemporal criterion map.

[0119] First, the stability prediction output generated by the digital twin model under the current shield tunneling conditions is input into the main discrimination path. The stratum stability response trend marked in the criterion map serves as the supervision reference path, and the two establish a correspondence in the spatiotemporal coordinates. Subsequently, structural consistency constraints are introduced into the adversarial learning mechanism to ensure that the residual extraction process focuses on local offsets or trend bifurcation errors rather than structural disturbances while keeping the topological structure of the sensitivity tensor unchanged, thereby avoiding topological instability caused by overfitting.

[0120] Secondly, the network employs a residual filter, extracting the difference signal between the model prediction and the criterion response through tensor error convolutional comparison operations. A graph structure preserver is introduced during this process to maintain the structural invariance of spatial dependencies and multi-physics coupling paths. The adversarial training module further optimizes the model's sensitivity to real-world perturbation trends, improving the robustness of the shift response extraction.

[0121] Finally, all significant residuals are encoded into incremental update vector groups, the structure of which includes timestamps, spatial locations, response types, and offset magnitudes, providing incremental basis for subsequent parameter updates and state corrections in the Siamese model.

[0122] This step significantly enhances the digital twin system's perception accuracy of stability shifts and its ability to maintain structure by introducing an adversarial residual extraction mechanism and structural consistency constraints. This method can monitor the dynamic error between the twin model output and the actual evolution on-site in real time under complex disturbance conditions, ensuring the accuracy and controllability of the update strategy. This lays a core foundation for achieving adaptive iteration of the twin system through "optimization during construction."

[0123] Step S64: Based on the joint optimization model of convolutional Markov state transition diagram and kernel projection, perform dynamic digital twin fusion processing on the incremental update vector group under constraints to ensure that the control response after model update is aligned with the original trajectory time, forming a time-consistent incremental twin model.

[0124] It can be understood that the joint optimization model of convolutional Markov state transition graph and kernel projection is a model that maps the incremental vector set to the Markov state space and introduces the kernel projection optimization mechanism. In the space of the joint optimization model of convolutional Markov state transition graph and kernel projection, different control response states (such as changes in tunneling speed, support pressure disturbance response, etc.) are defined as discrete Markov state nodes, and the disturbance response trend corresponding to the incremental vector serves as the input of potential state jump events. By introducing a temporal convolution module on the graph structure, the system can capture the evolution continuity and change trend of incremental data within a local time period, and enhance the temporal context awareness capability of state transition.

[0125] Subsequently, a kernel projection optimization mechanism is introduced to align the response trajectory of the original twin model with the current fitted trajectory triggered by incremental updates. By maximizing the similarity between the two in the kernel feature space (Gaussian kernel function), joint optimization based on temporal consistency constraints is completed. This process ensures that while absorbing new perturbation response trends, the model retains structural memory of historical control logic and formation response evolution paths, avoiding model behavior drift or response abrupt changes caused by updates.

[0126] The resulting incremental twin model possesses two key capabilities: firstly, it incorporates structural corrections caused by disturbance-sensitive points under the latest operating conditions, enhancing the model's adaptability to changes in risk conditions; secondly, it maintains alignment with the original control response path on the time axis, ensuring the continuity of the construction process and the stability of the control strategy.

[0127] Step S65: By constructing a non-homogeneous strategy clusterer based on the joint modeling of the jumping Markov model and the Dirichlet process Gaussian mixture model, the incrementally updated digital twin model is subjected to cross-time domain control map fusion processing to construct a set of locally optimal control paths with the goal of minimizing risk cost, and a dynamic management scheme for the entire shield tunneling construction process is obtained.

[0128] It is understandable that in order to achieve unified control and strategy selection for the digital twin model after incremental updates across multiple stages and working conditions, this step involves constructing a non-homogeneous strategy clusterer based on the joint modeling of a jumping Markov model and a Dirichlet process Gaussian mixture model. This clusterer is used to perform the fusion processing of cross-time domain control maps and generate dynamic management schemes that meet the risk control objectives of the entire construction process.

[0129] Specifically, a skip Markov model is first used to model the state transition trajectory in the incremental twin model, allowing states to jump between different time periods with non-constant probabilities, thereby characterizing irregular abrupt changes such as geological disturbances and construction strategy switching. This model is particularly suitable for capturing non-stationary state transitions caused by external disturbances (such as water inrush, weak layers, and stress release in surrounding rock) during construction, overcoming the limitation that traditional Markov chains only support stable transitions.

[0130] After the state sequence modeling is completed, the system introduces the Dirichlet Process Gaussian Mixture Model (DPGMM) as a state clusterer. Without a preset number of categories, it performs probabilistic clustering of control states formed under different time periods and geological conditions, thereby forming multiple "strategy state clusters" with internal response consistency. Each cluster represents the strategy performance under a stable working condition, such as a low-speed tunneling-high-pressure support combination in high-permeability strata, or a high-power propulsion-low-intervention mode in hard rock strata. By performing path optimization based on minimizing risk potential energy (such as collapse risk and water inrush intensity) within each strategy cluster, the system generates a corresponding set of locally optimal control paths. These paths are weighted by the response effects of control variables (velocity, pressure, slurry density) under different states, constituting the core structure of the dynamic control map of shield tunneling.

[0131] Ultimately, the paths corresponding to each strategy cluster are merged into a dynamic management scheme map that can span the time domain and adapt to multiple state transitions. This map not only supports strategy prediction for future time windows, but also allows for real-time adjustment of the current strategy to respond to emergencies.

[0132] Example 2:

[0133] like Figure 2 As shown, this embodiment provides a shield tunneling construction management device based on digital twins. See [link to documentation]. Figure 2 The device includes an acquisition unit 701, a first processing unit 702, a second processing unit 703, a third processing unit 704, a control unit 705, and an evaluation unit 706.

[0134] Acquisition unit 701 is used to acquire basic data for shield tunneling construction in karst formations;

[0135] The first processing unit 702 is used to process the basic data through a fusion process of multi-scale curvature spectrum clustering method and cross-domain spectrum graph generation method to obtain a spatiotemporal multimodal hypermap that synchronously represents the multi-source relationship of stratigraphy, hydrology and mechanism.

[0136] The second processing unit 703 is used to fuse the spatiotemporal multimodal hypergraph with a structure-preserving graph embedding algorithm and a graph neural network based on a diffusion process to obtain a dynamically coupled digital twin model.

[0137] The third processing unit 704 is used to process the digital twin model by combining persistent spectrum enhancement processing and random walk algorithm to obtain a dynamically evolving tunneling risk topology spectrum sequence.

[0138] The control unit 705 is used to adaptively control the shield tunneling speed, support pressure and mud parameters based on the tunneling risk topology spectrum sequence, and generate an adaptive tunneling control strategy.

[0139] The evaluation unit 706 is used to perform real-time stability evaluation based on the adaptive tunneling control strategy, and to incrementally update the digital twin model based on the real-time evaluation results, so as to obtain a dynamic management scheme for the entire shield tunneling construction process.

[0140] It should be noted that the specific manner in which each module performs its operation in the apparatus described in the above embodiments has been described in detail in the embodiments of the method, and will not be elaborated here.

[0141] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0142] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A shield tunneling construction management method based on digital twins, characterized in that, include: Obtain basic data for shield tunneling construction; The basic data is processed by a fusion of multi-scale curvature spectrum clustering method and cross-domain spectral graph generation method to obtain a spatiotemporal multimodal hypergraph that synchronously represents the multi-source relationship of stratigraphy, hydrology and mechanism. The spatiotemporal multimodal hypergraph is fused with a structure-preserving graph embedding algorithm and a diffusion-based graph neural network to obtain a dynamically coupled digital twin model. The digital twin model is then processed by a combination of persistent spectrum enhancement and random walk algorithm to obtain a dynamically evolving tunneling risk topology spectrum sequence. An adaptive tunneling control strategy is generated based on the adaptive regulation of shield tunneling speed, support pressure and mud parameters according to the tunneling risk topology spectrum sequence. Based on the aforementioned adaptive tunneling control strategy, a real-time stability assessment is performed, and the digital twin model is incrementally updated based on the real-time assessment results to obtain a dynamic management scheme for the entire shield tunneling construction process. The basic data is fused using a multi-scale curvature spectrum clustering method and a cross-domain spectral graph generation method, including: The basic data is subjected to multi-scale curvature spectrum clustering and curvature islands are adaptively divided according to fractal dimension to obtain primary clusters of stratigraphic geometry. The seasonal groundwater level time series in the primary cluster of the stratigraphic geometry and the basic data are subjected to cross-domain spectral resonance alignment to obtain a multi-source spectral resonance matrix. The multi-source spectral resonance matrix is ​​subjected to cross-domain spectral graph generation processing to obtain a priori set of supernodes containing three types of supernodes: virtual voids, water masses, and faults, and their weights. The prior hypernode set is reconstructed using a spatiotemporal hypergraph to obtain a spatiotemporal multimodal hypergraph that synchronously represents the multi-source relationship between stratigraphy, hydrology, and mechanism. The process of fusing the spatiotemporal multimodal hypergraph with a structure-preserving graph embedding algorithm and a diffusion-based graph neural network includes: Based on the spatiotemporal multimodal hypergraph, a high-order structure embedding process is performed to construct a structure-preserving embedding vector set, which is a vector set that preserves formation fracture modes, grouting behavior and seepage pressure disturbance in a high-dimensional embedding space. Based on the nonlinear isostructural transformation of the structure-preserving embedded vector set, the structural disturbance trajectory in the stratum attribute space and the feedback evolution path in the shield control variable space are cross-domain mapped to construct an isostructural nonlinear mapping model. Based on the isomorphic nonlinear mapping model, a graph neural network based on the diffusion process is used for modeling to construct a graph neural network that reflects the relationship between the stratum disturbance response mechanism and the shield control behavior. The dynamic twin mechanism is constructed based on the diffusion process-based graph neural network. By performing dynamic graph memory modeling on the diffusion structure of the diffusion process-based graph neural network at the time series level, a dynamically coupled digital twin model is constructed.

2. The shield tunneling construction management method based on digital twin as described in claim 1, characterized in that... The digital twin model is then subjected to a fusion process of persistent spectrum enhancement and random walk algorithm, including: On the time frame of the digital twin model, the supernodes of the spatiotemporal multimodal hypergraph are processed by spectral wavelet transform based on spectral graph theory, and combined with the Vietoris-Rips complex construction method, the spatiotemporal connection relationship between the supernodes is gradually mapped into a nested hierarchical simplex complex to obtain a cross-scale topological filter cluster. The topological filter cluster is subjected to persistent spectrum enhancement processing based on the method of fusion β distribution fitting and persistent entropy measurement to obtain a set of homogeneous persistent feature vectors. Based on the homology persistence feature vector set, random walk trend field encoding is performed to generate a node-edge risk potential energy matrix that reflects the sensitivity of water inrush-collapse coupling. Based on the node-edge risk potential energy matrix, time series clustering analysis is performed. By capturing the migration and aggregation trajectory of risk potential energy peaks in the time domain, a dynamically evolving tunneling risk topology spectrum sequence is obtained.

3. The shield tunneling construction management method based on digital twins according to claim 1, characterized in that... The adaptive control of shield tunneling speed, support pressure, and mud parameters based on the aforementioned tunneling risk topology spectrum sequence includes: Based on the tunneling risk topology spectrum sequence, counterfactual causal inference is performed. By injecting the changes in shield control quantity as exogenous do intervention into the hypergraph-Do Bayesian network, a multi-field coupled causal structure graph with scenario-based counterfactual weights is generated. The multi-field coupled causal structure diagram is processed by fractional manifold spectrum mapping according to the Caputo fractional manifold embedding method to obtain a measurable shield control-stratum response dynamic manifold embedding tensor. Based on the manifold embedding tensor, a variable step size Stackelberg difference game model is constructed on the strategy surface of tunneling speed-support pressure-slurry density, and the Pareto optimal response curve is solved to obtain the staged control candidate tensor. The Hamiltonian variational inequality is solved based on the control candidate tensor to obtain an adaptive tunneling control strategy that satisfies dynamic stability constraints.

4. A shield tunneling construction management device based on digital twin, characterized in that, include: The acquisition unit is used to acquire basic data for shield tunneling construction. The first processing unit is used to process the basic data through a fusion of a multi-scale curvature spectrum clustering method and a cross-domain spectral graph generation method to obtain a spatiotemporal multimodal hypergraph that synchronously represents the multi-source relationship between stratigraphy, hydrology, and mechanism. The second processing unit is used to fuse the spatiotemporal multimodal hypergraph with a structure-preserving graph embedding algorithm and a graph neural network based on a diffusion process to obtain a dynamically coupled digital twin model. The third processing unit is used to process the digital twin model by combining persistent spectrum enhancement processing and random walk algorithm to obtain a dynamically evolving tunneling risk topology spectrum sequence. The control unit is used to adaptively control the shield tunneling speed, support pressure and mud parameters based on the tunneling risk topology spectrum sequence, and generate an adaptive tunneling control strategy. The evaluation unit is used to perform real-time stability evaluation based on the adaptive tunneling control strategy, and to incrementally update the digital twin model based on the real-time evaluation results, so as to obtain a dynamic management scheme for the entire shield tunneling construction process. The first processing unit includes: The first processing subunit is used to perform multi-scale curvature spectrum clustering on the basic data and adaptively divide the curvature islands according to the fractal dimension to obtain the primary cluster of stratigraphic geometry. The second processing subunit is used to perform cross-domain spectral resonance alignment processing on the primary cluster of stratigraphic geometry and the seasonal groundwater level time series in the basic data to obtain a multi-source spectral resonance matrix. The third processing subunit is used to perform cross-domain spectral graph generation processing on the multi-source spectral resonance matrix to obtain a priori supernode set containing three types of supernodes: virtual void, water mass, and fault, and their weights. The fourth processing subunit is used to perform spatiotemporal supergraph reconstruction processing on the prior supernode set to obtain a spatiotemporal multimodal supergraph that synchronously represents the multi-source relationship between stratigraphy, hydrology, and mechanism. The second processing unit includes: The fifth processing subunit is used to perform high-order structure embedding processing based on the spatiotemporal multimodal hypergraph to construct a structure-preserving embedding vector set, which is a vector set that preserves formation fracture modes, grouting behavior and seepage pressure disturbance in a high-dimensional embedding space. The first construction sub-unit is used to perform nonlinear isostructural transformation processing based on the structure-preserving embedded vector set, and to perform cross-domain mapping between the structural disturbance trajectory in the stratum attribute space and the feedback evolution path in the shield control variable space to construct an isostructural nonlinear mapping model. The sixth processing subunit is used to perform modeling processing of a diffusion-based graph neural network based on the isomorphic nonlinear mapping model, and to construct a diffusion-based graph neural network that reflects the relationship between the stratum disturbance response mechanism and the shield control behavior. The second construction subunit is used to construct a dynamic twin mechanism based on a graph neural network based on a diffusion process. By performing dynamic graph memory modeling on the diffusion structure of the graph neural network based on the diffusion process at the time series level, a dynamically coupled digital twin model is constructed.

5. The shield tunneling construction management device based on digital twin according to claim 4, characterized in that, The third processing unit includes: The seventh processing subunit is used to perform spectral wavelet transform processing on the supernodes of the spatiotemporal multimodal hypergraph based on spectral graph theory on the time frame of the digital twin model, and combine the Vietoris-Rips complex construction method to gradually map the spatiotemporal connection relationship between the supernodes into a nested hierarchical simplex complex to obtain a cross-scale topological filter cluster. The eighth processing subunit is used to perform persistent spectrum enhancement processing on the topological filter cluster according to the method of fusion β distribution fitting and persistent entropy measurement to obtain a set of homogeneous persistent feature vectors; The ninth processing subunit is used to perform random walk trend field encoding processing based on the homology persistence feature vector set to generate a node-edge risk potential energy matrix that reflects the sensitivity of water inrush-collapse coupling. The tenth processing subunit is used to perform time series clustering analysis based on the node-edge risk potential energy matrix. By capturing the migration and aggregation trajectory of risk potential energy peaks in the time domain, a dynamically evolving tunneling risk topology spectrum sequence is obtained.

6. The shield tunneling construction management device based on digital twin according to claim 4, characterized in that, The control unit includes: The first control subunit is used to perform counterfactual causal inference processing based on the tunneling risk topology spectrum sequence. By injecting the shield control quantity change as an exogenous do intervention into the hypergraph-Do Bayesian network, a multi-field coupled causal structure graph with scenario-based counterfactual weights is generated. The second control subunit is used to perform fractional manifold spectrum mapping on the multi-field coupled causal structure diagram according to the Caputo fractional manifold embedding method to obtain a measurable shield control-stratum response dynamic manifold embedding tensor. The third control subunit is used to construct a variable step size Stackelberg difference game model on the strategy surface of tunneling speed-support pressure-slurry density based on the manifold embedding tensor, and solve the Pareto optimal response curve to obtain the staged control candidate tensor. The fourth control subunit is used to solve the Hamiltonian variational inequality based on the control candidate tensor to obtain an adaptive tunneling control strategy that satisfies dynamic stability constraints.

Images